VF-Nav: visual floor-plan-based point-goal navigation
In indoor navigation tasks, the use of floor plans is an efficient and cheap way to provide globally consistent metric and topological information about various environments. However, most studies on floor-plan-based navigation have relied on LiDAR rather than RGB cameras because of the difficulty of performing cross-modality matching. In this paper, we instead focus on the visual indoor navigation problem and propose VF-Nav, a visual floor-plan-based point-goal navigation algorithm combining a brain-inspired localization method with a topological planning technique. In the proposed approach,continuous and accurate localization is achieved by combining the metric information provided by the floor plan with a brain-inspired localization model. Then, the global path to the point goal is generated by building the topological map from the floor plan, and a short-term target is provided at each step. Finally, a reinforcement learning control module guides the robot to reach each short-term target. The experimental results on a simulated point-goal navigation dataset demonstrate the excellent performance of the proposed approach in a complicated indoor environment. Our method achieves a success rate of up to 88% and a success weighted by path length of 71%.
Science China(Information Sciences)
网络首发
立即查看 >
图书推荐
相关工具书